How I'd grow Amp

Amp is a coding agent. It runs in the terminal, in editors, in apps for Mac, iPhone and Android, and in orbs: cloud machines where every thread gets its own copy of your repository. It started inside Sourcegraph, and at the end of 2025 its founders spun it out as Amp Inc.

Amp makes money from plans. The free Hobby plan pays as you go for orbs. Megawatt costs $20 a month and includes 45,000 minutes of orb time, and Gigawatt is the bigger tier. Teams put members on those plans, and Enterprise is a sales conversation. Amp charges no token fees, and people can bring their own ChatGPT or Claude subscriptions.

So every visitor who matters is a developer choosing an agent, or a team choosing where its agents run.

Amp’s search footprint is mostly one domain. I checked the others: www.ampcode.com, ampcode.net and Sourcegraph’s old Amp address all redirect to ampcode.com, which is right. ampcode.app serves a full second copy of the site (more on that below). And Sourcegraph, now a separate company, still owns the pages for Cody, the assistant whose free and paid users it sent to Amp.

Almost all of Amp’s search traffic is the word “amp”

Ahrefs estimates ampcode.com gets about 3,300 visits from Google this month. This map shows where they land.

About 1% of those visits land on a page that sells something. 83% go to the homepage from people who typed Amp’s name and would have found it anyway. The biggest group after that is people searching for “orbs”, mostly for wheels, solar garden lights and Christmas decorations.

The traffic also looks like it collapsed. Ahrefs had ampcode.com at about 15,500 visits a month in February, and about 3,200 now.

The ranking didn’t drop. In February Ahrefs had Amp second for “amp”, which gets about 64,000 US searches a month, and credited it with about 2,950 visits from that word. Today Amp is first for the same word, and Ahrefs credits it with about 100. “Amp” also means amplifiers, amperes and Google’s AMP pages, so most people who type it probably want something else, and Ahrefs now estimates that most of them click something else.

These are third-party estimates, so the shape matters more than the exact numbers. Either way, none of this traffic was people shopping for a coding agent. Those people search for different words, and Amp has no pages for them.

Competitors with Amp’s authority get several times its traffic

These are Ahrefs’ estimates of this month’s Google traffic for the main coding agents.

OpenCode and Cursor are household names, so much of their traffic will be people typing their names. The useful comparison is Augment Code. Its site has the same Domain Rating by Ahrefs as Amp’s, and it gets about seven times the traffic. Kilo Code, one point higher, gets almost three times.

Neither got there with a bigger brand. They publish pages for the searches developers make before they’ve picked a tool.

Plans: buyers search for a coding agent, and Amp has no page for them

These are the searches people make when they’re choosing an agent, with US monthly volumes from Ahrefs.

“coding agent” has a difficulty of 1, and “claude code alternative” has 2. That’s close to zero. Both are searches from people who are about to pick, or switch, the tool they pay for. Amp’s Domain Rating of 76 matches Augment Code and Qodo, and both of them are on page one.

A Domain Rating 16 roundup is on page one. So is Augment Code, whose roundup pages also rank in the top three for “ai coding tools” and “ai coding assistant”. Amp publishes strong opinion pieces (its post “The coding agent is dead” is a good example), but none of them is aimed at a search.

Even the comparison with Amp’s own name in it belongs to other sites.

Someone searching “amp vs claude code” already knows Amp and is close to signing up. Right now a directory with a Domain Rating of 52 tells them the answer. Amp should.

I’d build three kinds of page. First, comparisons against Claude Code, Cursor, Codex and OpenCode that are honest about where each one wins. Second, alternatives pages for “claude code alternative” and “cursor alternative”, which also answer the questions in Google’s AI Overview for those searches. Third, one guide to what a coding agent is and how to choose one, written in Amp’s own voice, which is opinionated enough to stand out on a results page full of near-identical lists of the top ten tools. Google’s AI Overview for “claude code alternative” currently cites DigitalOcean, Kilo and Zapier. Being on the lists that rank is what gets Amp named in AI answers too.

Orbs: the page ranks for the wrong word

Orbs are what Megawatt and Gigawatt sell: minutes of cloud machine time for agents. The orbs page has about 8,900 words, a screencast and around 410 internal links. Its title is “What Are Orbs? - Amp”, so it ranks for “orbs”, and that brings in people looking for wheels and garden lights.

Developers who want what orbs do search for sandboxes and cloud environments.

“claude code sandbox” is the one I’d go after first. Below Anthropic’s two docs pages, it’s a Reddit thread, a Substack post and a Medium post. Amp lets people run Claude through their own subscription, and orbs are a sandbox for exactly that.

“ai sandbox” is mixed: some of it is universities offering staff a safe chatbot. But Northflank and CircleCI, two developer tool companies, rank with “what is an AI sandbox” pages, which tells me the developer meaning wins clicks too.

I’d keep “What Are Orbs?” as the heading, because it’s the product’s name and it’s a good line. I’d change the title tag Google shows, name sandboxes and cloud environments in the first paragraph, and add one page for running Claude Code or Codex in an orb.

The model database is ready to rank, and Google can’t find it

Amp has a database of 109 AI models, with context limits, output limits, tool support and release dates, and a page for each model. People compare models before they choose which one to pay for, and Amp sells access to all of them.

The demand is real. “gpt-6 astra” gets about 3,500 US searches a month at a difficulty of 0, and “ai model comparison” gets about 1,400 at 12.

OpenCode is a coding agent, and it ranks sixth with the kind of page Amp already has. Kilo’s leaderboard ranks second for “best ai model for coding” (900 searches) as well. Amp’s database doesn’t rank for either, and the section on technical fixes below shows why.

Amp’s model content does rank when Google can find it. Its GPT-5.5 model card, which is in the sitemap, ranks fourth for “gpt-5.5 xhigh”. I’d give every database page the same chance, then add a short “use it in Amp” block to each one: which mode runs it and what it costs. That turns a reference page into a page that sells.

Sourcegraph still gets the Cody searches

Amp grew out of Sourcegraph, which also makes Cody, a coding assistant. When Sourcegraph ended Cody Free and Pro, it pointed those users to Amp. The two companies are separate now, but Cody’s searches still land on Sourcegraph’s site.

I crawled 600 sourcegraph.com pages to see how they link to Amp. There are 15 links in all, and 8 of them point at Amp’s homepage with the link text “Amp”. The Cody docs page, which Ahrefs estimates gets about 2,800 visits a month, has none.

That line is Sourcegraph’s call, since it still sells Cody Enterprise. Its own post about the plan changes already sends Cody Pro users to Amp with $40 in credits, so it’s the same message on the page that gets the traffic.

Amp can do its part alone. “sourcegraph cody” gets about 600 US searches a month, and the sixth result is a Reddit thread titled “Replaced by Cody AMP is there Alternative?”. A page on moving from Cody to Amp would answer that thread’s question on Amp’s own site.

New pages: the searches Amp’s competitors win

Cursor, Claude Code, Augment Code, Kilo, OpenCode and Cline rank in Google’s top 20 for searches a person choosing a coding agent makes, and Amp has no page for any of them.

Most of these pages are help articles and docs. Cursor ranks first for “agentic coding” and tenth for “multi-agent” with help-centre pages, and the multi-agent one is 841 words. Claude Code’s worktrees page is a docs page. Kilo is first for “how to cancel github copilot subscription” with a how-to titled “Cancel GitHub Copilot Subscription: Steps & Alternative”. Amp has docs on most of these subjects (orbs, Puck, agent to agent, AGENTS.md), but none of them is written for a search.

The plans section above covers the pages with Amp’s own name in them. These are the four groups I’d add next.

Comparisons between the agents Amp runs

Amp can use a Claude subscription to run its Claude Code mode, and a ChatGPT one for GPT models. Its pricing docs say so. So Amp’s team works with both model families every day, and can run the same tasks in Codex and Claude Code on one repository and publish the threads as evidence.

The biggest comparison searches are between other agents, and neither agent being compared ranks for its own comparison.

Duet’s page is fifth with about 3,000 words, five tables and “Updated November 2026” at the top. It’s a blog post written by someone who doesn’t make either agent. “cursor vs claude code” looks the same: a personal blog with a Domain Rating of 42 is second.

I’d publish them as “Codex vs Claude Code: the same tasks, run in both” at /compare/codex-vs-claude-code and “Cursor vs Claude Code: an editor or a terminal agent” at /compare/cursor-vs-claude-code. Each one runs three real tasks, links the threads, and ends with how to use both in Amp.

What coding agents cost

“claude code pricing” gets about 20,000 US searches a month at a difficulty of 11. Below Reddit (“Claude Pro at $20 is a trap for dev work”) and Anthropic’s own cost docs, it’s Superblocks and Softr, two app-builder companies. Softr’s page has about 5,000 words and four tables.

I’d call it “What coding agents cost: Claude Code, Codex, Cursor and Amp” at /guides/coding-agent-pricing, with a visible updated date, because prices change and readers check how old a pricing page is. The last section is about Amp: bring the plan you already pay for, and Amp adds no token fees on it (on every tier but Enterprise).

Running agents in parallel

Orbs are Amp’s answer to “how do I run several agents at once”: every thread gets its own cloud machine with its own copy of the repository, and Puck launches and coordinates them. People search for the problem in git’s words.

“multi-agent” is mostly people asking about AI research (Wikipedia and LangChain’s docs top the results), so I’d count on the worktree searches more. The guide would be “How to run coding agents in parallel: worktrees, orbs and subagents” at /guides/parallel-coding-agents. It starts where the searcher is, with worktrees on one laptop, and shows where they stop working: shared ports, one database, one machine’s CPU.

Model pages and a free tool, built from the model database

The model database already holds what people search for when they compare models: context and output limits, reasoning levels, tool support, knowledge cutoff and release date for 109 models. The biggest pattern is one model against another.

OpenCode already does this. Its page comparing two versions of DeepSeek V4 Flash ranks first for “deepseek v4 flash 0731” (2,900 searches), and its HTML holds about 320 words. The rest draws in the browser. Amp’s pages would carry the full data in the HTML, say which Amp mode uses each model, and quote Amp’s own evals where it has a model card, as it does for GPT-5.5.

I’d build a pair page only when people search for that pair. Every pair of 109 models is 5,886 pages, and most would be thin pages nobody asked for. Start with the pairs above and add one each time a new model ships and a pair starts getting searches.

The free tool is a token counter.

“token counter” gets about 4,500 US searches a month, “token calculator” 1,200 and “claude token counter” 500 at a difficulty of 6. Amp’s version would count the tokens, then show which of the 109 models the text fits in and how much of each context window it fills. The counting can use OpenAI’s published tokenizer and Anthropic’s token counting endpoint, and the limits come from Amp’s model database.

llms.txt lists only the docs, and the pricing answers aren’t in the HTML

AI crawlers get the same full text a browser gets (eight pages, requested as GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot), so the gaps are in what that text leaves out.

The pricing page matters most, because it holds the answer to a question buyers ask AI assistants. “Are there any fees for using my own subscriptions or keys?” is on the page, and its answer isn’t in the HTML that any crawler reads.

The fix: render every answer in the HTML and fold it with a details element or CSS, so people still see a tidy list. Add FAQPage structured data that matches the visible text. Then give llms.txt a short product section above the docs index, with Markdown copies of the pricing, orbs and models pages.

Amp is a coding agent and development
environment built for the frontier.

This is an index of the Amp documentation
for agents and other programmatic readers.

## Get Started

- [Introduction](https://ampcode.com/docs/markdown)
  ...

Amp is a coding agent and development
environment built for the frontier.

## Product

- [Pricing](https://ampcode.com/pricing.md):
  Hobby is free, Megawatt is $20 a month
  with 45,000 orb minutes, Gigawatt is $200.
  No Amp token fees on your own ChatGPT or
  Claude subscriptions, except on Enterprise.
- [What are orbs?](https://ampcode.com/what-are-orbs.md):
  Cloud sandboxes where each agent thread
  gets its own copy of your repository.
- [Models](https://ampcode.com/models.md):
  Context, output limits and release dates
  for 109 models.
- [Chronicle](https://ampcode.com/chronicle):
  Release news, notes from the team,
  guides and podcast episodes.

## Docs

- [Introduction](https://ampcode.com/docs/markdown)
  ...

To confirm it worked, request the pricing page as GPTBot and check the answers are in the response, then request ampcode.com/pricing.md and check for Markdown instead of the sign-in page.

AI answers name Cursor in most answers and Amp in none

Amp wasn’t cited or named in any of 28 answers from ChatGPT search, Google AI Mode and Google’s AI Overviews to ten questions a developer asks before choosing an agent (seo research ai_answers, DataForSEO, US, 4 November 2026). Cursor, which Ahrefs credits with about 399,000 Google visits a month, was named in 17 of 27 answers to the same ten and cited in 4.

Two kinds of page get cited. The first is lists written by someone else. Zapier’s posts on AI coding tools, Claude Code alternatives and Cursor alternatives turned up in 5 of the 20 source lists, more often than Cursor’s own site. None of them mentions Amp, and neither do the lists from Builder.io, Faros and Built In.

The second is a page titled with the question. OpenAI’s help centre has “Using Codex with your ChatGPT plan”, and it was cited for the subscription question. Warp has a guide to running several coding agents, and it was cited for the parallel one.

Amp has its own answer to two of these questions. Its pricing page says “Use your ChatGPT, Claude & other subs” and “No Amp token fees or limits”. But for “Which AI coding agent has no token fees?”, ChatGPT cited four small sites and no Amp page. Amp has no page with the question as its title and a short answer at the top.

The fix:

1. Publish one page per question Amp has an answer to, with the question as the title and a two-sentence answer first: using your ChatGPT or Claude subscription in Amp, running Claude Code in the cloud in an orb, and the parallel agents guide above.
2. Ask to be added to the lists: Zapier’s three posts, Builder.io’s Cursor alternatives, Faros’s best coding agents and Built In’s comparison. Each one is an email to the author.

To confirm it worked, ask the same ten questions again after 30 days and count the answers that cite or name Amp.

Three technical fixes, closest to revenue first

The crawl found three problems I’d fix. I’ve ordered them by how close each one sits to revenue.

1. ampcode.app serves a second copy of the site

ampcode.app returns the same pages as ampcode.com with a 200 status instead of redirecting, from the homepage to the docs. Most of them carry a canonical tag that points back to ampcode.com, which limits the damage. The pricing page has no canonical tag on either domain, so Google sees two equal copies of the page that sells the plans. Ahrefs has no keywords for ampcode.app yet, so this is a fix before it costs anything.

The fix: a permanent (301) redirect from every ampcode.app address to the same path on ampcode.com, set up as one rule at the CDN. Then add a canonical tag to the pricing page (and every other page without one) that names its own ampcode.com address. To confirm it worked, request ampcode.app/pricing and check the response is a 301, then check the pricing page in Search Console’s URL Inspection, where the user-declared canonical should be the ampcode.com address.

2. The model database is missing from the sitemap

The sitemap lists 317 addresses. One of them is a model page, the GPT-5.5 model card. The database itself and all 109 model pages are missing, the site’s “Models” links go to the modes page instead, and the model pages I checked have no meta description or canonical tag. My crawl of 582 pages, which followed links from the homepage and the sitemap, found none of them.

The model card and the database page for GPT-5.5 also share one title tag, so Google has to pick between them. Every database page uses the same pattern, the model’s name and “Amp”, which says nothing about what’s on the page.

The fix: add the database and every model page to the sitemap, and link the database from the site menu and the footer. Give each model page a self canonical and a description written from its own data, such as its context limit and release date. Give the database page for GPT-5.5 a different title from the model card, or point it at the card. To confirm it worked, check that the sitemap lists about 110 more addresses, then watch Search Console’s page indexing report for the model pages over the next two weeks.

3. Unknown addresses redirect to the sign-in page

Ask ampcode.com for an address that doesn’t exist and it redirects to the sign-in page, which returns a 200 status. To Google, every mistyped or retired address looks like a working page. DataForSEO still shows Google ranking some of these, such as the address of a Neovim plugin file that now ends at “Sign in to Amp”.

GET /greggh/claude-code.nvim/…
  302 Found
  → /auth/sign-in?returnTo=…
  200 OK "Sign in to Amp"

GET /greggh/claude-code.nvim/…
  404 Not Found

GET /threads/T-…
  302 Found
  → /auth/sign-in?returnTo=…

Amp’s news section already gets this right: an unknown news address returns a 404. The fix is to do the same everywhere else.

The fix: return a 404 for any address that isn’t a real page or app route, and only redirect real app routes (such as threads and settings) to sign-in. For a retired address that had links or rankings, add a 301 to the nearest live page instead. While they’re in there, the docs link to the orbs profiles page is broken and needs pointing at the right page. To confirm it worked, request a made-up address and check for a 404, then check that Search Console’s soft 404 count stops growing.

Ship it all today, then let it settle

Teardowns like this usually end with a four-week plan. With coding agents, I’d start all of it on the same day, and Amp, of all companies, has the orbs to run it.

Each fix above is a separate job with a clear end, so these can run at once:

1. 1

   Comparison pages

   Amp vs Claude Code, Cursor, Codex and OpenCode, plus Claude Code and Cursor alternatives pages, honest about where each tool wins.

2. 2

   Coding agent guide

   One guide to what a coding agent is and how to choose one, linked from the homepage and the footer.

3. 3

   Orbs for search

   A new title tag on the orbs page, and a page for running Claude Code or Codex in an orb.

4. 4

   Model database

   Sitemap entries, a menu link, self canonicals and descriptions for all 109 model pages, plus a "use it in Amp" block on each.

5. 5

   One domain

   Redirect ampcode.app to ampcode.com and add a canonical tag to the pricing page.

6. 6

   Real 404s

   Return 404 for unknown addresses, keep sign-in redirects for app routes, and fix the orbs profiles docs link.

7. 7

   Moving from Cody

   A page for Cody Free and Pro users, and a note to Sourcegraph asking for one line on the Cody docs page.

8. 8

   Codex and Cursor comparisons

   Codex vs Claude Code and Cursor vs Claude Code, each with three real tasks run in both and the threads linked.

9. 9

   Pricing guide

   What coding agents cost, with Claude Code, Codex, Cursor and coding plans, and a dated last section on bringing your own plan to Amp.

10. 10

    Parallel agents guide

    Worktrees, orbs and subagents in one guide that answers how to run several agents at once.

11. 11

    Model pairs and token counter

    A page for each model pair people search for, and a token counter that shows which models the text fits.

12. 12

    AI readiness

    Pricing answers in the HTML, a product section in llms.txt, and Markdown copies of the product pages.

13. 13

    Answer pages

    One page each for using your ChatGPT or Claude subscription in Amp and running Claude Code in the cloud, titled with the question.

14. 14

    The lists

    Emails to the authors of the Zapier, Builder.io, Faros and Built In lists that AI answers cite and that leave Amp out.

Someone at Amp still needs to review the comparison pages and the guides, because they make claims about competitors. The redirects, titles, sitemap and canonical changes need only a quick check.

Then I’d spend 30 days checking the results:

1. Day 0

   Day 0: Ask Google to recrawl

   Submit every new and changed URL to Index My Site, so Google recrawls them now instead of on its own schedule.

2. Day 3

   Day 3: Check indexing

   Confirm the comparison pages and model pages are indexed, and that ampcode.app now redirects.

3. Day 7

   Day 7: First impressions

   Look for impressions on "claude code alternative", "coding agent" and "claude code sandbox" in Search Console.

4. Day 14

   Day 14: Model pages

   Check how many model pages have impressions, and which models people search for most.

5. Day 30

   Day 30: Sign-ups

   Compare sign-ups and paid plans that started from search with the month before.

6. Day 30

   Day 30: AI answers

   Ask ChatGPT, AI Mode and AI Overviews the same ten questions for Amp and for Cursor, and count the answers that cite or name Amp.

Get the changes seen quickly

The model pages are the biggest test here. Google has never been told about 109 of them, and a sitemap entry only asks it to visit when it gets round to it. I’d submit every new and changed URL to Index My Site the day it ships, so Google is asked to crawl it straight away. That matters most for the model pages, the retitled orbs page and the new comparison pages, where nothing can move until Google has seen them.

How I ran this

Everything in this teardown came from SEO Skill, run by an agent. You can run the same audit on your own site.

1. Connect the hosted MCP server. It’s free. You sign in with Google, and there’s no password and nothing to install. In Claude Code:

claude mcp add --transport http seoskill https://seoskill.dev/mcp

Codex, Cursor and Claude on the web work too. The MCP guide has the steps for each one.

2. Add a research key. On your account page, save an Ahrefs or DataForSEO key. The key gives your agent search volumes, rankings and results pages, like the tables above.

3. Install the CLI for crawling. The crawl runs on your machine, so it can read any site:

npm i -g seo
seo login

4. Ask your agent. This is close to the prompt I used:

Audit example.com and every domain we own.
Follow redirects from our old and sister domains.
Crawl each one with seo crawl, and check what
unknown addresses and the sitemap return.
Use seo_research to find the searches our buyers
make, who ranks for them and whether we can compete.
Pull our top competitors' ranking keywords and list
the ones we have no page for, grouped into pages.
Ask ai_answers ten buyer questions for us and for
our biggest competitor.
Rank every fix by how close it sits to revenue.

On your own site, the hosted server adds the part I couldn’t see from the outside: your real Search Console clicks and queries. With those, you get measured numbers where this post has estimates.

Check every domain that leads to you, too. Two of the fixes in this teardown came from addresses nobody looks at: a second domain and a hundred pages Google was never told about.

If you’re at Amp, I’m happy to talk any of this through. And if you’d like the same teardown for your company, SEO Skill will run it for you.